encodemb.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include "./av1_rtcd.h"
#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"

#include "aom_dsp/bitwriter.h"
#include "aom_dsp/quantize.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"

#include "av1/common/idct.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/scan.h"

#include "av1/encoder/av1_quantize.h"
#include "av1/encoder/encodemb.h"
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/tokenize.h"

#if CONFIG_PVQ
#include "av1/encoder/encint.h"
#include "av1/common/partition.h"
#include "av1/encoder/pvq_encoder.h"
#endif

#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif

// Check if one needs to use c version subtraction.
static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; }

static void subtract_block(const MACROBLOCKD *xd, int rows, int cols,
                           int16_t *diff, ptrdiff_t diff_stride,
                           const uint8_t *src8, ptrdiff_t src_stride,
                           const uint8_t *pred8, ptrdiff_t pred_stride) {
#if !CONFIG_HIGHBITDEPTH
  (void)xd;
#endif

  if (check_subtract_block_size(rows, cols)) {
#if CONFIG_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8,
                                  src_stride, pred8, pred_stride, xd->bd);
      return;
    }
#endif  // CONFIG_HIGHBITDEPTH
    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                         pred_stride);

    return;
  }

#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride,
                              pred8, pred_stride, xd->bd);
    return;
  }
#endif  // CONFIG_HIGHBITDEPTH
  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                     pred_stride);
}

void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
                      int blk_col, int blk_row, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const int diff_stride = block_size_wide[plane_bsize];
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  const int tx1d_width = tx_size_wide[tx_size];
  const int tx1d_height = tx_size_high[tx_size];
  uint8_t *dst =
      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
  uint8_t *src =
      &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  int16_t *src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
  subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src,
                 src_stride, dst, dst_stride);
}

void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];
  const MACROBLOCKD *xd = &x->e_mbd;

  subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
                 pd->dst.buf, pd->dst.stride);
}

// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
#if CONFIG_EC_ADAPT
  { 10, 7 }, { 8, 5 },
#else
  { 10, 6 }, { 8, 6 },
#endif
};

#define UPDATE_RD_COST()                             \
  {                                                  \
    rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0); \
    rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1); \
  }

static INLINE unsigned int get_token_bit_costs(
    unsigned int token_costs[2][COEFF_CONTEXTS][ENTROPY_TOKENS], int skip_eob,
    int ctx, int token) {
#if CONFIG_NEW_TOKENSET
  (void)skip_eob;
  return token_costs[token == ZERO_TOKEN || token == EOB_TOKEN][ctx][token];
#else
  return token_costs[skip_eob][ctx][token];
#endif
}

#define USE_GREEDY_OPTIMIZE_B 0

#if USE_GREEDY_OPTIMIZE_B

typedef struct av1_token_state_greedy {
  int16_t token;
  tran_low_t qc;
  tran_low_t dqc;
} av1_token_state_greedy;

static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
                             int block, TX_SIZE tx_size, int ctx) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const int ref = is_inter_block(&xd->mi[0]->mbmi);
  av1_token_state_greedy tokens[MAX_TX_SQUARE + 1][2];
  uint8_t token_cache[MAX_TX_SQUARE];
  const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  const int eob = p->eobs[block];
  const PLANE_TYPE plane_type = pd->plane_type;
  const int16_t *const dequant_ptr = pd->dequant;
  const uint8_t *const band_translate = get_band_translate(tx_size);
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
  int dqv;
  const int shift = av1_get_tx_scale(tx_size);
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
#elif !CONFIG_AOM_QM
  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
#endif  // CONFIG_NEW_QUANT
  int sz = 0;
  const int64_t rddiv = mb->rddiv;
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
  int i, final_eob;
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  const int default_eob = tx_size_2d[tx_size];

  assert(mb->qindex > 0);

  assert((!plane_type && !plane) || (plane_type && plane));
  assert(eob <= default_eob);

  int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;

  int64_t rate0, rate1;
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    int x = qcoeff[rc];
    t0 = av1_get_token(x);

    tokens[i][0].qc = x;
    tokens[i][0].token = t0;
    tokens[i][0].dqc = dqcoeff[rc];

    token_cache[rc] = av1_pt_energy_class[t0];
  }
  tokens[eob][0].token = EOB_TOKEN;
  tokens[eob][0].qc = 0;
  tokens[eob][0].dqc = 0;
  tokens[eob][1] = tokens[eob][0];

  unsigned int(*token_costs_ptr)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      token_costs;

  final_eob = 0;

  int64_t eob_cost0, eob_cost1;

  const int ctx0 = ctx;
  /* Record the r-d cost */
  int64_t accu_rate = 0;
  int64_t accu_error = 0;

  rate0 = get_token_bit_costs(*(token_costs_ptr + band_translate[0]), 0, ctx0,
                              EOB_TOKEN);
  int64_t best_block_rd_cost = RDCOST(rdmult, rddiv, rate0, accu_error);

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;

  int x_prev = 1;

  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    int x = qcoeff[rc];
    sz = -(x < 0);

    int band_cur = band_translate[i];
    int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
    int token_tree_sel_cur = (x_prev == 0);

    if (x == 0) {
      // no need to search when x == 0
      rate0 =
          get_token_bit_costs(*(token_costs_ptr + band_cur), token_tree_sel_cur,
                              ctx_cur, tokens[i][0].token);
      accu_rate += rate0;
      x_prev = 0;
      // accu_error does not change when x==0
    } else {
      /*  Computing distortion
       */
      // compute the distortion for the first candidate
      // and the distortion for quantizing to 0.
      int dx0 = (-coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx0 >>= xd->bd - 8;
      }
#endif
      int64_t d0 = (int64_t)dx0 * dx0;

      int x_a = x - 2 * sz - 1;
      int64_t d2, d2_a;

      int dx;

#if CONFIG_AOM_QM
      int iwt = iqmatrix[rc];
      dqv = dequant_ptr[rc != 0];
      dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
#else
      dqv = dequant_ptr[rc != 0];
#endif

      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx >>= xd->bd - 8;
      }
#endif  // CONFIG_HIGHBITDEPTH
      d2 = (int64_t)dx * dx;

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
      if (x_a != 0) {
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
             (coeff[rc] << shift);
#if CONFIG_HIGHBITDEPTH
        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
          dx >>= xd->bd - 8;
        }
#endif  // CONFIG_HIGHBITDEPTH
#else   // CONFIG_NEW_QUANT
#if CONFIG_HIGHBITDEPTH
        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
          dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
        } else {
          dx -= (dqv + sz) ^ sz;
        }
#else
        dx -= (dqv + sz) ^ sz;
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
      } else {
        d2_a = d0;
      }
      /*  Computing rates and r-d cost
       */

      int best_x, best_eob_x;
      int64_t base_bits, next_bits0, next_bits1;
      int64_t next_eob_bits0, next_eob_bits1;

      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
      rate0 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t0);

      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
      rate1 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t1);

      next_bits0 = 0;
      next_bits1 = 0;
      next_eob_bits0 = 0;
      next_eob_bits1 = 0;

      if (i < default_eob - 1) {
        int ctx_next, token_tree_sel_next;
        int band_next = band_translate[i + 1];

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x == 0);

        next_bits0 = get_token_bit_costs(*(token_costs_ptr + band_next),
                                         token_tree_sel_next, ctx_next,
                                         tokens[i + 1][0].token);
        next_eob_bits0 =
            get_token_bit_costs(*(token_costs_ptr + band_next),
                                token_tree_sel_next, ctx_next, EOB_TOKEN);

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x_a == 0);

        next_bits1 = get_token_bit_costs(*(token_costs_ptr + band_next),
                                         token_tree_sel_next, ctx_next,
                                         tokens[i + 1][0].token);

        if (x_a != 0) {
          next_eob_bits1 =
              get_token_bit_costs(*(token_costs_ptr + band_next),
                                  token_tree_sel_next, ctx_next, EOB_TOKEN);
        }
      }

      rd_cost0 = RDCOST(rdmult, rddiv, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, rddiv, (rate1 + next_bits1), d2_a);

      best_x = (rd_cost1 < rd_cost0);

      eob_cost0 = RDCOST(rdmult, rddiv, (accu_rate + rate0 + next_eob_bits0),
                         (accu_error + d2 - d0));
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
        eob_cost1 = RDCOST(rdmult, rddiv, (accu_rate + rate1 + next_eob_bits1),
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

      int dqc, dqc_a = 0;

      dqc = dqcoeff[rc];
      if (best_x + best_eob_x) {
        if (x_a != 0) {
#if CONFIG_NEW_QUANT
          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
                                           dequant_val[band_translate[i]]);
          dqc_a = shift ? ROUND_POWER_OF_TWO(dqc_a, shift) : dqc_a;
          if (sz) dqc_a = -dqc_a;
#else
// The 32x32 transform coefficient uses half quantization step size.
// Account for the rounding difference in the dequantized coefficeint
// value when the quantization index is dropped from an even number
// to an odd number.

#if CONFIG_AOM_QM
          tran_low_t offset = dqv >> shift;
#else
          tran_low_t offset = dq_step[rc != 0];
#endif
          if (shift & x_a) offset += (dqv & 0x01);

          if (sz == 0)
            dqc_a = dqcoeff[rc] - offset;
          else
            dqc_a = dqcoeff[rc] + offset;
#endif  // CONFIG_NEW_QUANT
        } else {
          dqc_a = 0;
        }  // if (x_a != 0)
      }

      // record the better quantized value
      if (best_x) {
        qcoeff[rc] = x_a;
        dqcoeff[rc] = dqc_a;

        accu_rate += rate1;
        accu_error += d2_a - d0;
        assert(d2_a <= d0);

        token_cache[rc] = av1_pt_energy_class[t1];
      } else {
        accu_rate += rate0;
        accu_error += d2 - d0;
        assert(d2 <= d0);

        token_cache[rc] = av1_pt_energy_class[t0];
      }

      x_prev = qcoeff[rc];

      // determine whether to move the eob position to i+1
      int64_t best_eob_cost_i = eob_cost0;

      tokens[i][1].token = t0;
      tokens[i][1].qc = x;
      tokens[i][1].dqc = dqc;

      if ((x_a != 0) && (best_eob_x)) {
        best_eob_cost_i = eob_cost1;

        tokens[i][1].token = t1;
        tokens[i][1].qc = x_a;
        tokens[i][1].dqc = dqc_a;
      }

      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
    assert(tokens[final_eob - 1][1].qc != 0);
    i = final_eob - 1;
    int rc = scan[i];
    qcoeff[rc] = tokens[i][1].qc;
    dqcoeff[rc] = tokens[i][1].dqc;
  }

  for (i = final_eob; i < eob; i++) {
    int rc = scan[i];
    qcoeff[rc] = 0;
    dqcoeff[rc] = 0;
  }

  mb->plane[plane].eobs[block] = final_eob;
  return final_eob;
}

#else  // USE_GREEDY_OPTIMIZE_B

typedef struct av1_token_state_org {
  int64_t error;
  int rate;
  int16_t next;
  int16_t token;
  tran_low_t qc;
  tran_low_t dqc;
  uint8_t best_index;
} av1_token_state_org;

static int optimize_b_org(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
                          int block, TX_SIZE tx_size, int ctx) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const int ref = is_inter_block(&xd->mi[0]->mbmi);
  av1_token_state_org tokens[MAX_TX_SQUARE + 1][2];
  uint8_t token_cache[MAX_TX_SQUARE];
  const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  const int eob = p->eobs[block];
  const PLANE_TYPE plane_type = pd->plane_type;
  const int default_eob = tx_size_2d[tx_size];
  const int16_t *const dequant_ptr = pd->dequant;
  const uint8_t *const band_translate = get_band_translate(tx_size);
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
  int dqv;
  const int shift = av1_get_tx_scale(tx_size);
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
#elif !CONFIG_AOM_QM
  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
#endif  // CONFIG_NEW_QUANT
  int next = eob, sz = 0;
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
  const int64_t rddiv = mb->rddiv;
  int64_t rd_cost0, rd_cost1;
  int rate0, rate1;
  int64_t error0, error1;
  int16_t t0, t1;
  int best, band = (eob < default_eob) ? band_translate[eob]
                                       : band_translate[eob - 1];
  int pt, i, final_eob;
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  const uint16_t *band_counts = &band_count_table[tx_size][band];
  uint16_t band_left = eob - band_cum_count_table[tx_size][band] + 1;
  int shortcut = 0;
  int next_shortcut = 0;

#if CONFIG_EXT_DELTA_Q
  const int qindex = cm->seg.enabled
                         ? av1_get_qindex(&cm->seg, xd->mi[0]->mbmi.segment_id,
                                          cm->base_qindex)
                         : cm->base_qindex;
  assert(qindex > 0);
  (void)qindex;
#else
  assert(mb->qindex > 0);
#endif

  token_costs += band;

  assert((!plane_type && !plane) || (plane_type && plane));
  assert(eob <= default_eob);

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
  tokens[eob][0].next = default_eob;
  tokens[eob][0].token = EOB_TOKEN;
  tokens[eob][0].qc = 0;
  tokens[eob][1] = tokens[eob][0];

  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    tokens[i][0].rate = av1_get_token_cost(qcoeff[rc], &t0, cat6_bits);
    tokens[i][0].token = t0;
    token_cache[rc] = av1_pt_energy_class[t0];
  }

  for (i = eob; i-- > 0;) {
    int base_bits, dx;
    int64_t d2;
    const int rc = scan[i];
    int x = qcoeff[rc];
#if CONFIG_AOM_QM
    int iwt = iqmatrix[rc];
    dqv = dequant_ptr[rc != 0];
    dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
#else
    dqv = dequant_ptr[rc != 0];
#endif
    next_shortcut = shortcut;

    /* Only add a trellis state for non-zero coefficients. */
    if (UNLIKELY(x)) {
      error0 = tokens[next][0].error;
      error1 = tokens[next][1].error;
      /* Evaluate the first possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

      if (next_shortcut) {
        /* Consider both possible successor states. */
        if (next < default_eob) {
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 +=
              get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
          rate1 +=
              get_token_bit_costs(*token_costs, 0, pt, tokens[next][1].token);
        }
        UPDATE_RD_COST();
        /* And pick the best. */
        best = rd_cost1 < rd_cost0;
      } else {
        if (next < default_eob) {
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 +=
              get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
        }
        best = 0;
      }

      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx >>= xd->bd - 8;
      }
#endif  // CONFIG_HIGHBITDEPTH
      d2 = (int64_t)dx * dx;
      tokens[i][0].rate += (best ? rate1 : rate0);
      tokens[i][0].error = d2 + (best ? error1 : error0);
      tokens[i][0].next = next;
      tokens[i][0].qc = x;
      tokens[i][0].dqc = dqcoeff[rc];
      tokens[i][0].best_index = best;

      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

      // The threshold of 3 is empirically obtained.
      if (UNLIKELY(abs(x) > 3)) {
        shortcut = 0;
      } else {
#if CONFIG_NEW_QUANT
        shortcut = ((av1_dequant_abscoeff_nuq(abs(x), dqv,
                                              dequant_val[band_translate[i]]) >
                     (abs(coeff[rc]) << shift)) &&
                    (av1_dequant_abscoeff_nuq(abs(x) - 1, dqv,
                                              dequant_val[band_translate[i]]) <
                     (abs(coeff[rc]) << shift)));
#else  // CONFIG_NEW_QUANT
#if CONFIG_AOM_QM
        if ((abs(x) * dequant_ptr[rc != 0] * iwt >
             ((abs(coeff[rc]) << shift) << AOM_QM_BITS)) &&
            (abs(x) * dequant_ptr[rc != 0] * iwt <
             (((abs(coeff[rc]) << shift) + dequant_ptr[rc != 0])
              << AOM_QM_BITS)))
#else
        if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
            (abs(x) * dequant_ptr[rc != 0] <
             (abs(coeff[rc]) << shift) + dequant_ptr[rc != 0]))
#endif  // CONFIG_AOM_QM
          shortcut = 1;
        else
          shortcut = 0;
#endif  // CONFIG_NEW_QUANT
      }

      if (shortcut) {
        sz = -(x < 0);
        x -= 2 * sz + 1;
      } else {
        tokens[i][1] = tokens[i][0];
        next = i;

        if (UNLIKELY(!(--band_left))) {
          --band_counts;
          band_left = *band_counts;
          --token_costs;
        }
        continue;
      }

      /* Consider both possible successor states. */
      if (!x) {
        /* If we reduced this coefficient to zero, check to see if
         *  we need to move the EOB back here.
         */
        t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
        t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
        base_bits = 0;
      } else {
        base_bits = av1_get_token_cost(x, &t0, cat6_bits);
        t1 = t0;
      }

      if (next_shortcut) {
        if (LIKELY(next < default_eob)) {
          if (t0 != EOB_TOKEN) {
            token_cache[rc] = av1_pt_energy_class[t0];
            pt = get_coef_context(nb, token_cache, i + 1);
            rate0 += get_token_bit_costs(*token_costs, !x, pt,
                                         tokens[next][0].token);
          }
          if (t1 != EOB_TOKEN) {
            token_cache[rc] = av1_pt_energy_class[t1];
            pt = get_coef_context(nb, token_cache, i + 1);
            rate1 += get_token_bit_costs(*token_costs, !x, pt,
                                         tokens[next][1].token);
          }
        }

        UPDATE_RD_COST();
        /* And pick the best. */
        best = rd_cost1 < rd_cost0;
      } else {
        // The two states in next stage are identical.
        if (next < default_eob && t0 != EOB_TOKEN) {
          token_cache[rc] = av1_pt_energy_class[t0];
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 +=
              get_token_bit_costs(*token_costs, !x, pt, tokens[next][0].token);
        }
        best = 0;
      }

#if CONFIG_NEW_QUANT
      dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
           (coeff[rc] << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx >>= xd->bd - 8;
      }
#endif  // CONFIG_HIGHBITDEPTH
#else   // CONFIG_NEW_QUANT
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
      } else {
        dx -= (dqv + sz) ^ sz;
      }
#else
      dx -= (dqv + sz) ^ sz;
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_NEW_QUANT
      d2 = (int64_t)dx * dx;

      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
      tokens[i][1].error = d2 + (best ? error1 : error0);
      tokens[i][1].next = next;
      tokens[i][1].token = best ? t1 : t0;
      tokens[i][1].qc = x;

      if (x) {
#if CONFIG_NEW_QUANT
        tokens[i][1].dqc = av1_dequant_abscoeff_nuq(
            abs(x), dqv, dequant_val[band_translate[i]]);
        tokens[i][1].dqc = shift ? ROUND_POWER_OF_TWO(tokens[i][1].dqc, shift)
                                 : tokens[i][1].dqc;
        if (sz) tokens[i][1].dqc = -tokens[i][1].dqc;
#else
// The 32x32 transform coefficient uses half quantization step size.
// Account for the rounding difference in the dequantized coefficeint
// value when the quantization index is dropped from an even number
// to an odd number.

#if CONFIG_AOM_QM
        tran_low_t offset = dqv >> shift;
#else
        tran_low_t offset = dq_step[rc != 0];
#endif
        if (shift & x) offset += (dqv & 0x01);

        if (sz == 0)
          tokens[i][1].dqc = dqcoeff[rc] - offset;
        else
          tokens[i][1].dqc = dqcoeff[rc] + offset;
#endif  // CONFIG_NEW_QUANT
      } else {
        tokens[i][1].dqc = 0;
      }

      tokens[i][1].best_index = best;
      /* Finally, make this the new head of the trellis. */
      next = i;
    } else {
      /* There's no choice to make for a zero coefficient, so we don't
       *  add a new trellis node, but we do need to update the costs.
       */
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      pt = get_coef_context(nb, token_cache, i + 1);
      /* Update the cost of each path if we're past the EOB token. */
      if (t0 != EOB_TOKEN) {
        tokens[next][0].rate += get_token_bit_costs(*token_costs, 1, pt, t0);
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != EOB_TOKEN) {
        tokens[next][1].rate += get_token_bit_costs(*token_costs, 1, pt, t1);
        tokens[next][1].token = ZERO_TOKEN;
      }
      tokens[i][0].best_index = tokens[i][1].best_index = 0;
      shortcut = (tokens[next][0].rate != tokens[next][1].rate);
      /* Don't update next, because we didn't add a new node. */
    }

    if (UNLIKELY(!(--band_left))) {
      --band_counts;
      band_left = *band_counts;
      --token_costs;
    }
  }

  /* Now pick the best path through the whole trellis. */
  rate0 = tokens[next][0].rate;
  rate1 = tokens[next][1].rate;
  error0 = tokens[next][0].error;
  error1 = tokens[next][1].error;
  t0 = tokens[next][0].token;
  t1 = tokens[next][1].token;
  rate0 += get_token_bit_costs(*token_costs, 0, ctx, t0);
  rate1 += get_token_bit_costs(*token_costs, 0, ctx, t1);
  UPDATE_RD_COST();
  best = rd_cost1 < rd_cost0;

  final_eob = -1;

  for (i = next; i < eob; i = next) {
    const int x = tokens[i][best].qc;
    const int rc = scan[i];
    if (x) final_eob = i;
    qcoeff[rc] = x;
    dqcoeff[rc] = tokens[i][best].dqc;

    next = tokens[i][best].next;
    best = tokens[i][best].best_index;
  }
  final_eob++;

  mb->plane[plane].eobs[block] = final_eob;
  assert(final_eob <= default_eob);
  return final_eob;
}

#endif  // USE_GREEDY_OPTIMIZE_B

int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
                   TX_SIZE tx_size, int ctx) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  const int eob = p->eobs[block];
  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
  if (eob == 0) return eob;
  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return eob;
#if CONFIG_PVQ
  (void)cm;
  (void)tx_size;
  (void)ctx;
  return eob;
#endif

#if USE_GREEDY_OPTIMIZE_B
  return optimize_b_greedy(cm, mb, plane, block, tx_size, ctx);
#else   // USE_GREEDY_OPTIMIZE_B
  return optimize_b_org(cm, mb, plane, block, tx_size, ctx);
#endif  // USE_GREEDY_OPTIMIZE_B
}

#if !CONFIG_PVQ
#if CONFIG_HIGHBITDEPTH
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
  QUANT_FUNC_HIGHBD = 1,
  QUANT_FUNC_TYPES = 2
} QUANT_FUNC;

static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
#if !CONFIG_NEW_QUANT
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
#else   // !CONFIG_NEW_QUANT
      { av1_quantize_fp_nuq_facade, av1_highbd_quantize_fp_nuq_facade },
      { av1_quantize_b_nuq_facade, av1_highbd_quantize_b_nuq_facade },
      { av1_quantize_dc_nuq_facade, av1_highbd_quantize_dc_nuq_facade },
#endif  // !CONFIG_NEW_QUANT
      { NULL, NULL }
    };

#else

typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
  QUANT_FUNC_TYPES = 1
} QUANT_FUNC;

static AV1_QUANT_FACADE quant_func_list[AV1_XFORM_QUANT_TYPES]
                                       [QUANT_FUNC_TYPES] = {
#if !CONFIG_NEW_QUANT
                                         { av1_quantize_fp_facade },
                                         { av1_quantize_b_facade },
                                         { av1_quantize_dc_facade },
#else   // !CONFIG_NEW_QUANT
                                         { av1_quantize_fp_nuq_facade },
                                         { av1_quantize_b_nuq_facade },
                                         { av1_quantize_dc_nuq_facade },
#endif  // !CONFIG_NEW_QUANT
                                         { NULL }
                                       };
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_PVQ

void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
#if !(CONFIG_PVQ || CONFIG_DAALA_DIST)
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
  PLANE_TYPE plane_type = get_plane_type(plane);
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const int is_inter = is_inter_block(mbmi);
  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, is_inter);
  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint16_t *const eob = &p->eobs[block];
  const int diff_stride = block_size_wide[plane_bsize];
#if CONFIG_AOM_QM
  int seg_id = mbmi->segment_id;
  const qm_val_t *qmatrix = pd->seg_qmatrix[seg_id][!is_inter][tx_size];
  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!is_inter][tx_size];
#endif

  FWD_TXFM_PARAM fwd_txfm_param;

#if CONFIG_PVQ || CONFIG_DAALA_DIST
  uint8_t *dst;
  int16_t *pred;
  const int dst_stride = pd->dst.stride;
  int tx_blk_size;
  int i, j;
#endif

#if !CONFIG_PVQ
  const int tx2d_size = tx_size_2d[tx_size];
  QUANT_PARAM qparam;
  const int16_t *src_diff;

  src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
  qparam.log_scale = av1_get_tx_scale(tx_size);
#if CONFIG_NEW_QUANT
  qparam.tx_size = tx_size;
  qparam.dq = get_dq_profile_from_ctx(x->qindex, ctx, is_inter, plane_type);
#endif  // CONFIG_NEW_QUANT
#if CONFIG_AOM_QM
  qparam.qmatrix = qmatrix;
  qparam.iqmatrix = iqmatrix;
#endif  // CONFIG_AOM_QM
#else
  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
  int skip = 1;
  PVQ_INFO *pvq_info = NULL;
  uint8_t *src;
  int16_t *src_int16;
  const int src_stride = p->src.stride;

  (void)ctx;
  (void)scan_order;
  (void)qcoeff;

  if (x->pvq_coded) {
    assert(block < MAX_PVQ_BLOCKS_IN_SB);
    pvq_info = &x->pvq[block][plane];
  }
  src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  src_int16 =
      &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];

  // transform block size in pixels
  tx_blk_size = tx_size_wide[tx_size];
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++)
        src_int16[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
  } else {
#endif  // CONFIG_HIGHBITDEPTH
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++)
        src_int16[diff_stride * j + i] = src[src_stride * j + i];
#if CONFIG_HIGHBITDEPTH
  }
#endif  // CONFIG_HIGHBITDEPTH
#endif

#if CONFIG_PVQ || CONFIG_DAALA_DIST
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
  pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];

  // transform block size in pixels